Tunnel construction deformation monitoring system based on machine vision and monitoring method thereof

Abstract

The invention relates to the technical field of tunnel construction monitoring, and discloses a tunnel construction deformation monitoring system based on machine vision and a monitoring method thereof. According to the method, real-time image acquisition is carried out on a tunnel construction structure to obtain a tunnel surface image sequence, and meanwhile, displacement parameters and strain parameters of the tunnel structure are monitored; performing initial deformation prediction according to the displacement parameter and the strain parameter to obtain a predicted deformation parameter; acquiring key area images in the tunnel by means of machine vision equipment, and inputting the key area images into a plurality of deformation feature recognizers respectively by combining the predicted deformation parameters to recognize and obtain actual deformation features; and finally, final deformation trend analysis is performed in combination with the predicted deformation parameters and the actual deformation characteristics, and a tunnel deformation monitoring result is obtained. According to the method, pertinence, comprehensiveness and accuracy of deformation monitoring are improved, the deformation state of the tunnel structure can be reflected more accurately, and an effective technical means is provided for safety monitoring of tunnel construction.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel construction monitoring, and in particular to a machine vision-based tunnel construction deformation monitoring system and a monitoring method thereof. Background Art

[0002] During tunnel construction, structural deformation monitoring is a crucial step in ensuring construction safety and project quality. Tunnel projects operate in a complex and dynamic geological environment, subject to multiple factors, including stratum pressure, groundwater action, and construction disturbances. These factors can easily lead to surrounding rock deformation, support structure cracking, and settlement. Failure to promptly detect and provide early warning can result in serious accidents such as collapse, resulting in significant casualties and economic losses.

[0003] Traditional tunnel deformation monitoring methods primarily rely on manual inspections and contact sensor monitoring. Manual inspections require personnel to enter the tunnel and perform data measurements using instruments such as total stations and levels. This is not only time-consuming and labor-intensive, but also subject to the harsh tunnel environment (such as dust, insufficient lighting, and cramped space). This results in low measurement efficiency, long data collection cycles, and difficulty achieving real-time monitoring. Monitoring lags often exist, preventing the timely capture of dynamic deformation changes. Furthermore, the accuracy of manual measurements is easily affected by factors such as operator skill and subjective judgment, resulting in significant data errors.

[0004] While contact sensor monitoring methods, such as strain gauges and displacement meters, can achieve a certain degree of automated monitoring, they have significant limitations. First, the sensors require direct contact with the tunnel structure, potentially damaging the support structure and compromising its integrity. Second, the limited coverage of the sensors, limited by their number and placement, makes it difficult to fully reflect deformation across the entire tunnel, potentially leaving blind spots for localized deformation in key areas. Furthermore, factors such as vibration and electromagnetic interference in the tunnel construction environment can easily lead to sensor failure or data drift, compromising the reliability of monitoring data. Furthermore, the high cost of maintaining and replacing sensors adds to the financial burden of engineering monitoring.

[0005] With the development of computer vision technology, machine vision monitoring methods have gradually been applied to tunnel deformation monitoring. Existing machine vision methods often collect tunnel surface images and use image processing algorithms to perform deformation analysis. However, these methods often rely solely on image information and lack consideration of the mechanical parameters of the tunnel structure. Because tunnel deformation is the result of the combined effects of structural forces and geological conditions, single image feature analysis is difficult to accurately reflect the inherent deformation mechanism and is easily affected by factors such as lighting changes, image noise, and surface texture differences. This results in low deformation recognition accuracy and makes it difficult to meet the engineering requirements for monitoring accuracy and reliability. Therefore, how to integrate multi-source monitoring data to improve the real-time, comprehensiveness, and accuracy of tunnel construction deformation monitoring has become a pressing issue in the field of tunnel engineering monitoring. Summary of the Invention

[0006] The purpose of the present invention is to provide a tunnel construction deformation monitoring method based on machine vision to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides a method for monitoring tunnel construction deformation based on machine vision, the method comprising:

[0008] Real-time image acquisition of tunnel construction structures is performed to obtain tunnel surface image sequences while monitoring the displacement and strain parameters of the tunnel structure.

[0009] According to the displacement parameters and strain parameters, the initial deformation is predicted to obtain the predicted deformation parameters;

[0010] Using machine vision equipment, images of key areas inside the tunnel are collected. Based on the predicted deformation parameters, these images are input into multiple deformation feature identifiers to identify the actual deformation features.

[0011] Combining the predicted deformation parameters with the actual deformation characteristics, the final deformation trend analysis is performed to obtain the tunnel deformation monitoring results.

[0012] Preferably, real-time image acquisition is performed on the tunnel construction structure to obtain a tunnel surface image sequence, while monitoring the displacement parameters and strain parameters of the tunnel structure, including:

[0013] Deploy multiple cameras at key points in the tunnel to continuously capture tunnel surface images and form a tunnel surface image sequence;

[0014] Use sensor networks to measure real-time displacement changes of tunnel structures and obtain displacement parameters;

[0015] Collect the strain data of the tunnel material through the strain gauge to obtain the strain parameters;

[0016] The displacement and strain parameters are associated with the timestamps of the tunnel surface image sequence.

[0017] Preferably, performing initial deformation prediction based on the displacement parameters and strain parameters to obtain predicted deformation parameters includes:

[0018] Collect historical tunnel construction data, extract sample displacement parameter sets and sample strain parameter sets, and mark the sample deformation dimensions to form a sample deformation parameter set;

[0019] Build a deformation prediction model based on time series analysis;

[0020] Using the sample displacement parameter set, sample strain parameter set, and sample deformation parameter set as training data and test data, the deformation prediction model is trained and tested, and the model optimization is completed after the error rate meets the standard;

[0021] Inputting displacement parameters and strain parameters into the deformation prediction model, and obtaining predicted deformation parameters through prediction output;

[0022] Compare the predicted deformation parameters with the real-time displacement parameters to calibrate the model deviation;

[0023] Based on the calibration model deviation, the deformation prediction model parameters are updated.

[0024] Preferably, a machine vision device is used to collect images of key areas inside the tunnel. Based on the predicted deformation parameters, the images of the key areas are respectively input into a plurality of deformation feature identifiers to identify and obtain actual deformation features, including:

[0025] Use machine vision equipment to focus on high-stress areas in the tunnel and capture images of key areas;

[0026] According to the predicted deformation parameters, the matching deformation level range is screened;

[0027] selecting a plurality of deformation feature identifiers corresponding to a deformation level range, each deformation feature identifier including a plurality of feature recognition paths based on image pattern matching;

[0028] The key area image is input into multiple deformation feature identifiers, and each deformation feature identifier outputs a binary recognition result;

[0029] Count the proportion of binary recognition results that are yes and calculate the probability distribution of deformation level;

[0030] The deformation level with the highest probability is selected as the actual deformation feature;

[0031] According to the actual deformation characteristics, feedback is provided to adjust the acquisition angle of the key area image.

[0032] Preferably, a plurality of deformation feature identifiers corresponding to the deformation level range are selected, each deformation feature identifier including a plurality of feature recognition paths based on image pattern matching, including:

[0033] Pre-training multiple deformation feature identifiers corresponding to multiple deformation levels, the training data including sample key area images and sample deformation feature binary results;

[0034] determining the number of feature recognition paths based on the error margin of the predicted deformation parameters;

[0035] randomly activating feature recognition paths within a plurality of deformable feature recognizers;

[0036] Input the key area image into the activated feature recognition path and output a set of binary recognition results;

[0037] Aggregate the binary recognition result set and calculate the actual deformation features.

[0038] Preferably, the predicted deformation parameters and actual deformation characteristics are combined to perform a final deformation trend analysis to obtain tunnel deformation monitoring results, including:

[0039] Integrate the predicted deformation parameters and actual deformation characteristics to form a deformation data vector;

[0040] Based on the moving window technique, the deformed data vector is time-sliced;

[0041] Analyze the deformation fluctuation characteristics within the time slice and calculate the average change rate of the deformation fluctuation characteristics;

[0042] Construct deformation trend indicators and generate tunnel deformation monitoring results based on the deformation trend indicators;

[0043] Compare tunnel deformation monitoring results with historical data to verify accuracy.

[0044] Preferably, the deformed data vector is time-sliced ​​based on a moving window technique, including setting a fixed time window length and a sliding step size, and dividing the deformed data vector into time series.

[0045] Preferably, constructing a deformation trend indicator includes:

[0046] Analyze the changing direction of deformation data within the time slice and calculate the upward and downward trend amounts of deformation data;

[0047] Based on the uptrend volume and the downtrend volume, the deformed trend indicator is obtained.

[0048] Preferably, the method further comprises: formatting a report of the tunnel deformation monitoring result, transmitting the report to a central monitoring system, and triggering an alarm mechanism based on the tunnel deformation monitoring result.

[0049] Preferably, a tunnel construction deformation monitoring system based on machine vision is characterized in that it is used to implement the above-mentioned tunnel construction deformation monitoring method based on machine vision, and the system includes:

[0050] Image acquisition module, used to collect real-time images of tunnel construction structures, obtain tunnel surface image sequences, and simultaneously monitor the displacement and strain parameters of the tunnel structure;

[0051] The prediction module is used to perform initial deformation prediction based on displacement parameters and strain parameters to obtain predicted deformation parameters;

[0052] The feature recognition module is used to collect images of key areas inside the tunnel through machine vision equipment. Based on the predicted deformation parameters, the images of key areas are input into multiple deformation feature identifiers to identify and obtain actual deformation features.

[0053] Analysis module, used to combine predicted deformation parameters and actual deformation characteristics to conduct final deformation trend analysis and obtain tunnel deformation monitoring results

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] This machine vision-based tunnel construction deformation monitoring method effectively overcomes the shortcomings of traditional monitoring methods through multi-dimensional data acquisition and fusion analysis. During data acquisition, tunnel surface image sequences are collected simultaneously with displacement and strain parameter monitoring, breaking the limitations of a single data source. Image sequences can intuitively reflect morphological changes on the tunnel surface, while displacement and strain parameters reveal the internal stress and deformation characteristics of the structure from a mechanical perspective. This combination provides richer basic data for deformation analysis, avoiding the incomplete information that comes from relying solely on images or single mechanical parameters.

[0056] The initial deformation prediction phase, based on displacement and strain parameters, provides a preliminary assessment of deformation trends from the perspective of structural mechanical response, providing clear guidance for subsequent image acquisition and analysis of key areas. This prediction method, based on mechanical parameters, makes image acquisition by machine vision equipment more targeted, avoiding the waste of computing power and information redundancy caused by indiscriminate image acquisition of the entire tunnel, thereby improving monitoring efficiency. By focusing on key areas, areas likely to experience significant deformation can be more accurately captured, reducing the interference of invalid data on monitoring results.

[0057] During the actual deformation feature recognition process, images of key areas are fed into multiple deformation feature identifiers, combined with predicted deformation parameters. This leverages the advantages of multi-identifier collaborative analysis. Different deformation feature identifiers are specifically designed to identify different manifestations of tunnel deformation (such as crack propagation, surface settlement, and support structure displacement), comprehensively capturing various deformation characteristics in key areas. This segmented recognition approach avoids the potential for single identifiers to miss or misjudge complex deformation features, thereby improving the completeness and accuracy of actual deformation feature extraction.

[0058] The final deformation trend analysis achieved a deep fusion of mechanical prediction and visual observation by combining predicted deformation parameters and actual deformation characteristics. The predicted deformation parameters provide the possible trend of deformation from a theoretical level, while the actual deformation characteristics verify and supplement the specific manifestations of deformation from an observational level. The two confirm and complement each other, and can more comprehensively reflect the deformation state of the tunnel structure. This fusion analysis method not only takes into account the mechanical response law of the structure, but also combines the actual observed surface morphological changes, making the deformation monitoring results more in line with the actual project, and can more accurately judge the safety status of the tunnel structure, providing a more reliable basis for construction decisions. At the same time, the entire monitoring process realizes the automated connection of data collection and analysis, reduces manual intervention, improves the real-time and continuity of monitoring, and can promptly detect potential deformation risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a working principle diagram of the tunnel construction deformation monitoring method based on machine vision according to the present invention;

[0060] Figure 2 Flowchart for the refinement of initial deformation prediction;

[0061] Figure 3 Flowchart for the refinement of deformation feature identifier and path selection;

[0062] Figure 4 Flowchart refined for the final deformation trend analysis. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0064] See also Figure 1 The present invention provides a tunnel construction deformation monitoring system and a monitoring method based on machine vision, the method comprising:

[0065] A camera array deployed at key points in the tunnel captures real-time images of the tunnel surface. Structural displacement parameters and material strain parameters are simultaneously collected using a network of displacement sensors and strain gauges. The image sequence and sensor data are spatially and temporally aligned using a unified timestamp. A deformation prediction model trained on historical construction data analyzes the current displacement and strain parameters and outputs predicted deformation parameters. Machine vision equipment dynamically adjusts the camera angle based on the predicted results, focusing on high-stress areas to capture images of key regions. Multiple deformation feature identifiers operating in parallel perform multi-path feature analysis on the images and output actual deformation characteristics. A data fusion module combines the predicted parameters and actual characteristics to generate a deformation data vector. The final monitoring results are then output after moving window analysis and trend indicator calculation. The monitoring results automatically generate standardized reports and trigger a tiered early warning mechanism, achieving a closed-loop process from data acquisition to decision support.

[0066] Example 1: See Figure 2 The tunnel surface image acquisition system adopts a distributed deployment approach, with weather-resistant industrial cameras installed on the tunnel vault, side walls, and invert arch. These cameras feature IP67 protection, built-in anti-fog heating modules, and are equipped with 2-megapixel CMOS sensors and f / 1.8 large aperture lenses. Installation locations are determined based on geological radar scanning results, with a focus on fault zones, densely jointed areas, and support structure joints. The spacing between adjacent cameras is controlled within 3-5 meters to provide overlapping field of view. Images are acquired at a rate of 10 frames per second, with a resolution of 1920×1080 pixels. Data is compressed using H.265 encoding and transmitted via a fiber optic ring network. A simultaneously deployed displacement monitoring system utilizes Bragg grating sensing technology, with 48-core sensor fibers embedded in the primary support concrete. Measurement points are set every 2 meters along the tunnel's longitudinal direction. A fiber optic interrogator acquires data at a 100Hz sampling rate and uses a phase demodulation algorithm to achieve three-dimensional displacement measurement, achieving an axial accuracy of 0.02mm and a radial accuracy of 0.05mm. Strain monitoring utilizes full-bridge resistance strain gauges, with 120Ω foil strain gauges welded to key nodes of the steel arch to form a Wheatstone bridge circuit. The data acquisition module, equipped with a temperature compensation unit, collects microstrain data at a 50Hz frequency, with a range of ±3000με and a resolution of 1με. All sensor data is transmitted via a 5G industrial router using MQTT over TLS. Each data packet is timestamped with a GPS-accurate time stamp accurate to the millisecond. After receiving the data, the central processing system uses a time series alignment algorithm to establish a mapping between the image frames and the sensor data, keeping the alignment error within 20 milliseconds.

[0067] The deformation prediction model construction process consists of two phases: offline training and online prediction. Historical data collection covers the complete construction cycles of at least three similar tunnel projects. The data samples include displacement time series data, strain distribution data, and corresponding laser scanning deformation measurements. During data preprocessing, cubic spline interpolation is used to fill missing values, and Z-score normalization is used to eliminate dimensional differences. The model architecture uses a bidirectional LSTM neural network. The input layer has 12 neurons corresponding to six displacement components (axial and radial) and six strain partition means. The hidden layer consists of three layers: 64 LSTM units in the first layer, 32 GRU units in the second layer, and 16 fully connected nodes in the third layer. The output layer has three neurons, which output the predicted axial deformation, radial convergence value, and crown settlement. The training process uses a sliding window strategy with a window length of 120 minutes and a sliding step size of 10 minutes. The Nadam algorithm is used as the optimizer, and the loss function is a weighted combination of smoothed L1 loss and cosine similarity. The model undergoes full training every 24 hours, retaining dynamic samples from the last 30 days. During online prediction, real-time sensor data is filtered using a sliding mean and fed into the model, generating a set of prediction parameters every 30 seconds. The system compares the predicted displacement values ​​with the actual fiber optic sensor values ​​in real time. When the absolute deviation of five consecutive sampling points exceeds 0.15mm, the model's incremental update mechanism is triggered. This update utilizes an elastic weight hardening algorithm, adjusting the weights of the fully connected layers while preserving important parameters. Updates are completed within 200 milliseconds.

[0068] The displacement parameter acquisition system utilizes distributed fiber-optic acoustic sensing technology, with single-mode sensing fibers laid longitudinally along the tunnel. A demodulator transmits a 1550nm frequency-modulated continuous wave (FMCW) signal and uses coherent detection to analyze phase variations in the Rayleigh scattering signal. Data processing utilizes the φ-OTDR algorithm, with a spatial resolution of 1 meter and a measurement range covering a 2000-meter tunnel section. The system performs baseline calibration every 5 minutes to mitigate the effects of temperature drift. The strain monitoring network utilizes a hybrid network of vibrating-wire and resistive strain gauges, with vibrating-wire sensors deployed in areas of concrete stress concentration and resistive strain gauges installed at key points of rebar stress. The data acquisition terminal is equipped with a 6-channel, 24-bit ADC with a configurable sampling rate of 10-100Hz. The sensor network utilizes a star topology, with regional aggregation nodes connected to a central server via Industrial Ethernet. The spatiotemporal alignment system utilizes an improved dynamic time warping algorithm to establish temporal correspondences between image sequences and sensor data. The alignment process first extracts image time codes and sensor UTC timestamps, then uses linear interpolation to compensate for transmission delays, ultimately creating a spatiotemporal index table with millisecond-level accuracy.

[0069] The model training dataset contains three sets of data: displacement parameters, strain parameters, and corresponding total station-measured deformations. The data augmentation phase uses time warping and Gaussian noise injection to expand the sample size to five times the original data. Network training adopts a phased strategy: in the first phase, the LSTM layer weights are frozen, and only the fully connected layers are trained; in the second phase, all parameters are unfrozen and fine-tuned. Regularization measures include setting the weight decay coefficient to 1e-4 and applying dropout with a 50% probability to the hidden layers. The model is deployed using the TensorRT-optimized inference engine and runs on a Jetson AGX Xavier edge computing device. The prediction service is exposed via a gRPC interface, supporting 100 concurrent calls per second. The online calibration module sets two-level deviation thresholds: when the instantaneous deviation exceeds 0.3mm or the average deviation over 10 consecutive times exceeds 0.1mm, a model retraining task is automatically generated. The retraining process uses a transfer learning strategy, preserving the convolutional layer weights and updating only the fully connected layer parameters. Training is completed within 15 minutes. The calibrated model gradually replaces the online service through the canary release strategy to ensure the uninterrupted operation of the prediction service.

[0070] The system implements multi-source data fusion processing, with the image acquisition unit and sensor network achieving microsecond-level time synchronization via the PTPv2 protocol. The displacement data processing process includes three steps: outlier filtering (based on the 3σ criterion), trend removal (EMD decomposition), and downsampling (Kalman filtering). A spatial interpolation algorithm is used to generate a two-dimensional strain contour map for strain data, and key eigenvalues ​​are extracted using contour lines. The prediction model input vector is constructed using a sliding window method with a 60-minute window length, containing 720 displacement samples and 3,600 strain samples. During the feature engineering phase, 12 statistical features are extracted: mean, variance, range, and autocorrelation coefficient for displacement data; kurtosis, skewness, energy entropy, and power spectral density for strain data. Model output post-processing uses exponentially weighted moving average smoothing with a smoothing coefficient α set to 0.2. The deviation detection module uses a CUSUM control chart algorithm with a cumulative deviation threshold of 0.8 mm. When a model update is triggered, the system automatically creates a training task queue, prioritizing data from periods with high deviation. After training, A / B testing is performed to verify the effectiveness of the model improvements. The entire data processing process runs in a containerized environment, and elastic scheduling of computing resources is achieved through Kubernetes.

[0071] Example 2: See Figure 3Image capture of critical areas utilizes a high-precision pan-tilt camera system equipped with a three-axis servo motor drive, enabling ±180° pan rotation and +90° / -30° tilt adjustment. The pan-tilt camera is equipped with a 20-megapixel global shutter CMOS sensor and an f / 2.8-16 motorized zoom lens, achieving a resolution of 50μm / pixel at a minimum focus distance of 0.3 meters. After receiving the predicted deformation parameters, the system uses a spatial coordinate conversion algorithm to determine the coordinates of the target area's center point. The localization process uses an iterative closest point algorithm to match the predicted deformation position with the actual 3D point cloud model, maintaining a positioning error within 2 cm. When the predicted value enters the Level I (0-3mm), Level II (3-6mm), or Level III (>6mm) warning range, the pan-tilt camera automatically adjusts to the preset observation position. For Level III warning areas, the system activates macro mode, automatically switching the lens focal length to 150mm, and a ring-shaped LED fill light provides shadowless illumination.

[0072] The deformation feature recognition system consists of nine independent recognizers, with three groups corresponding to each deformation level. The recognizers utilize a containerized deployment architecture, and each recognizer group includes three feature processing paths. The first path uses a modified U-Net convolutional neural network, downsampling and upsampling the input image five times to extract multi-scale features. The second path applies a multi-directional Sobel operator for edge enhancement, combined with a Hough transform to detect linear features. The third path extracts texture features using a fusion algorithm of local binary patterns and gray-level co-occurrence matrices. Each path outputs a binary recognition result, which is then aggregated and judged using a weighted voting mechanism. The weight distribution is dynamically adjusted based on the path's historical accuracy, with an initial weighting of 0.4:0.3:0.3.

[0073] A two-stage transfer learning strategy was used for discriminator training. The base model used a ResNet-101 architecture pre-trained on the COCO dataset, with the input layer resized to 640×480 pixels to accommodate tunnel image scale. Fine-tuning training utilized a labeled dataset containing typical deformation features such as cracks, spalling, and leakage. The annotations included defect type, geometric dimensions, and spatial distribution. Data augmentation employed random rotation (±15°), brightness adjustment (±20%), and local occlusion. Training utilized layer-wise learning rate settings, with convolutional layers set to 1e-5 and fully connected layers to 1e-4. A focal loss loss was used to address sample imbalance, and the RAdam algorithm was used as the optimizer.

[0074] During the online recognition phase, the system activates the corresponding cluster of recognizers based on the predicted deformation level. Each recognizer uses a random path selection mechanism, randomly enabling two feature paths for each processing step. The outputs of the six activated paths are fed into a decision fusion module, which calculates the confidence level of crack identification and deformation estimates. When the confidence level falls below the 85% threshold, a multi-view recapture mechanism is triggered. This recapture strategy generates three supplementary observation angles: a primary view with a horizontal offset of ±15° and a +10° elevation view. Additional images are fed into the recognition system for cross-validation, and the final result is the median of the multi-view recognition results.

[0075] The image processing pipeline consists of three steps: preprocessing, feature extraction, and decision fusion. The preprocessing stage performs lens distortion correction and perspective transformation, converting oblique images into orthographic projections. In the feature extraction stage, the convolutional neural network path outputs a 1024-dimensional feature vector, the edge detection path generates an edge density distribution map, and the texture analysis path calculates texture response values ​​in eight directions. The decision fusion module uses a fuzzy inference system, with input parameters including crack continuity index, edge density peak, and texture disorder. Outputs include deformation grade classification and crack width estimation with millimeter-level accuracy.

[0076] The system incorporates a dynamic feedback mechanism to adjust acquisition parameters based on recognition results. When a Level III deformation feature is detected, the sampling rate is automatically increased to 30 frames per second and structured light projection mode is activated to acquire the three-dimensional shape. If the recognition confidence level remains below the threshold, the system automatically switches to a higher-resolution mode and increases the fill light intensity. All recognition results are associated with spatial coordinate information and displayed in real-time within the tunnel BIM model. The data processing pipeline utilizes an asynchronous architecture, with image acquisition, feature extraction, and result fusion executed in parallel across three computing nodes, keeping the single-frame processing latency within 300 milliseconds.

[0077] The deformation feature database utilizes a hierarchical storage structure, retaining original images for 30 days, feature vectors for 180 days, and recognition results permanently. The query interface supports searches based on pile number range, deformation level, and time interval. The system performs self-diagnosis every 24 hours, verifying the accuracy of the recognizer using standard test images. Any deviation exceeding 5% automatically triggers a model retraining process. This retraining process utilizes an online learning strategy, incrementally updating the parameters of the fully connected layers while maintaining the weights of the convolutional layers. The updated model is validated by running in shadow mode and, once verified, deployed to the production environment.

[0078] The pan / tilt control system utilizes a closed-loop feedback mechanism, with position accuracy verified via a rotary encoder and visual odometry. The motion control algorithm employs adaptive PID regulation, dynamically adjusting the proportional coefficient based on target distance. The anti-shake mechanism utilizes electronic image stabilization and mechanical vibration reduction to maintain image clarity even in vibrating construction environments. The lighting system features eight independently controllable LED arrays with adjustable color temperatures between 3000K and 6500K, automatically adjusting output power based on ambient light intensity. All device parameters are mapped in real time via the digital twin system, allowing operators to remotely adjust the shooting parameters of any camera from the control center.

[0079] Example 3: See Figure 4 Deformation data fusion processing employs a hierarchical and progressive architecture, integrating predicted deformation parameters with actual deformation features into a unified representation. The data vector construction process first performs principal component analysis on the displacement parameters, extracting the first three principal components as the displacement components. After the strain parameters are interpolated using Kriging to generate a two-dimensional distribution field, the gradient modulus in eight directions is calculated as the strain component. Visual features are compressed into a 256-dimensional feature vector using a convolutional neural network, and then dimensionality reduction is performed using t-SNE to obtain three visual components. The resulting 14-dimensional data vector is represented as V=[d1,d2,d3,s1,...,s8,v1,v2,v3], where d represents the principal component of displacement, s represents the strain gradient, and v corresponds to the visual feature. Vector elements are scaled to the [0,1] interval using min-max normalization to eliminate dimensionality differences.

[0080] Time window processing uses a double buffering mechanism, with a fixed 60-second main window and a 10-second sliding subwindow. The main window buffer adopts a circular queue structure with a capacity of 360 sampling points (corresponding to a 6Hz sampling rate). As the window slides, new data overwrites the oldest data to maintain time series continuity. The subwindow processing unit performs a discrete wavelet transform on the data within the window, using the db4 wavelet basis function for a five-layer decomposition, extracting approximate coefficients and detail coefficients as time-frequency features. The fluctuation characteristics of the displacement component are obtained by calculating the mean of the absolute value of the first-order difference, expressed as follows:

[0081]

[0082] in: represents the displacement fluctuation intensity, is the number of sampling points in the window, Represents the principal component of displacement at the i-th sampling point. Strain component fluctuation analysis uses the directional gradient method to calculate the Euclidean distance of strain gradients between adjacent sampling points. The visual component uses an optical flow algorithm to extract feature point motion vectors, and the histogram entropy of the vector amplitude is used as a fluctuation indicator.

[0083] The deformation trend analysis module employs a multi-scale feature fusion strategy. The linear regression slope and intercept of each component are calculated at a 60-second window scale. At a 10-minute macroscale, the trend and period terms are decomposed using a Hodrick-Prescott filter. The trend indicator construction process first performs a Hilbert transform on the displacement component to determine the instantaneous frequency and phase. The strain component obtains the energy change rate by calculating the time derivative of the strain energy density. A three-dimensional LBP operator is used to extract spatiotemporal texture features from the visual component. These features are weighted and fused using an attention mechanism, with the attention weights adaptively adjusted based on the feature variance.

[0084] The risk scoring model is implemented using a three-layer feedforward neural network. The input layer receives 14 feature parameters, the hidden layer has 32 neurons, and the output layer generates a risk score ranging from 0 to 100. The network training uses the Wasserstein distance as the loss function, and the optimization process applies a gradient penalty strategy. The model is automatically updated every 24 hours, and the training data retains the last seven days of labeled samples. The output score is filtered through a sliding average filter with a window width of 5 sampling points and a smoothing coefficient of 0.3.

[0085] The anomaly detection subsystem uses the isolation forest algorithm to construct 100 isolation trees in the feature space. Each tree is constructed by randomly selecting features and split values. The anomaly score is calculated as the average path length required to isolate a data point. When the anomaly score exceeds a dynamic threshold, a detailed diagnostic process is triggered. The diagnostic process includes three steps: displacement-strain cross-validation, multi-period data comparison, and spatial correlation analysis. The validation results generate an anomaly report, noting the possible cause and impact range.

[0086] The data visualization system maps the analysis results into a 3D heat map, which is overlaid on the tunnel BIM model. The heat map color coding uses the HSL color space, with hue indicating deformation type, saturation reflecting the rate of change, and brightness corresponding to the risk level. The view supports timeline dragging, allowing users to review deformation states at any moment. Interactive features include profile analysis tools, contour generators, and trend comparison views. All visualization elements are rendered using WebGL acceleration, supporting real-time, simultaneous viewing on multiple devices.

[0087] In terms of system implementation, the data processing pipeline adopts a microservices architecture, comprising five service units: data access, window calculation, feature extraction, trend analysis, and result output. Services communicate with each other via the gRPC protocol, with message serialization using the Protocol Buffers format. Compute-intensive tasks are deployed and run on GPU nodes, and the in-memory database uses a Redis cluster to store real-time data. The task scheduler is built on Kubernetes, supporting elastic scaling of computing resources. A quality monitoring module continuously tracks processing latency for each service, automatically triggering horizontal scaling when P99 latency exceeds 500 milliseconds.

[0088] The deformation feature database uses a time-partitioned storage strategy, with daily data stored independently in a columnar database. Zstandard is used as the compression algorithm, balancing compression ratio and decompression speed. The query interface supports SQL-like syntax and provides multi-dimensional filtering by spatial region, time range, and feature value. Data export formats include CSV, JSON, and Parquet to meet the needs of various analysis tools. All data operations are audit-logged, and blockchain technology ensures that the logs cannot be tampered with.

[0089] System integration testing uses fault injection to simulate abnormal scenarios such as network latency, data loss, and compute node failure. Recovery mechanisms include data retransmission, compute rollback, and service degradation. Performance testing shows that the system maintains an end-to-end processing delay of less than 800 milliseconds under an input load of 1,000 data points per second. Resource monitoring shows that under normal operating conditions, CPU utilization remains between 40% and 60%, and memory usage does not exceed 70% of the allocated amount. The security protection system includes three lines of defense: transport layer encryption, access control lists, and behavioral auditing. Penetration testing is performed regularly to update protection rules.

[0090] Example 4: The deformation trend analysis adopts a dual time window processing mechanism, with the main window width set to 300 seconds and the sub-window width to 60 seconds. The system slides the sub-window every 10 seconds and refreshes the main window data every 300 seconds. In the actual monitoring of a certain tunnel section, the system recorded data changes for 30 consecutive minutes. The displacement component processing flow first performs a linear fit on the displacement sequence in the main window and calculates the slope angle of the fitting line. When continuous deformation occurs at the monitoring pile number K25+380, the system captures the change process of the displacement slope gradually increasing from -0.5° to +2.3°. A positive slope angle indicates an upward trend, a negative value indicates a downward trend, and the absolute value reflects the intensity of the change. At the same time, the determination coefficient is calculated to evaluate the reliability of the linear fit, and when the coefficient is lower than 0.6, the nonlinear trend analysis is started.

[0091] Strain component analysis uses an energy density calculation method. The system integrates the strain gradients in eight directions within each subwindow to determine the strain energy density. Energy density values ​​in adjacent subwindows are compared using a relative rate of change calculation: the absolute difference between the current and previous values ​​divided by the previous value. When the energy density in the K25+380 section increased from 1.8 J / m³ to 2.7 J / m³ between 10:15 and 10:20, the system calculated a 50% rate of change. When this rate of change exceeds the 25% threshold, a warning flag is triggered.

[0092] Visual texture feature extraction utilizes an improved LBP-TOP algorithm. The system treats image sequences within 60-second subwindows as space-time cubes, calculating local binary patterns in the XY plane and dynamic texture changes in the X-T and Y-T planes. Each subwindow outputs a texture chaos index, ranging from 0 to 1, with larger values ​​indicating more dramatic surface texture changes. At 10:15 a.m. in the K25+380 segment, the system detected a jump in the texture chaos index of the vault concrete surface from 0.35 to 0.68, indicating abnormal surface deformation.

[0093] The trend indicator fusion is represented by a three-dimensional vector: [displacement trend angle, strain change rate, texture disorder]. The trend indicator changes at five consecutive time points in this segment are shown in Table 1.

[0094] Table 1: The changes in trend indicators for 5 consecutive time points in this segment are as follows.

[0095]

[0096] Risk value mapping is implemented using a radial basis function network. The network input layer receives three trend indicators, and the hidden layer uses 24 Gaussian kernel functions, with the center point determined by clustering historical data. The output layer uses linear combination to generate a risk value on a scale of 0-100. For the data in the table above, the system calculated a risk value of 73.5 at 10:15:20, exceeding the orange warning threshold of 60. Network parameters are automatically updated weekly, and training data uses labeled samples from the last three months.

[0097] The upward trend is calculated using a modified Mann-Kendall test. The system assesses trend significance within the displacement series by counting the percentage of sampling points that meet the following criteria: three consecutive points showing a monotonically increasing trend with a change exceeding the measurement error. In the K25+380 case, 83% of the sampling points detected during the period 10:15-10:18 met the upward trend criteria. The downward trend is determined through strain energy density analysis, calculating the average negative gradient of the energy density curve. A significant downward trend is determined when this value remains below -0.15 J / m³·s for three consecutive minutes.

[0098] In terms of system implementation, the trend analysis engine is deployed on a dedicated computing node equipped with dual Xeon processors and 128GB of memory. The data processing pipeline consists of four stages: data normalization, which scales raw parameters to a uniform dimension; feature extraction, which parallelizes the calculation of three trend indicators; fusion, which constructs input vectors; and assessment, which outputs risk values. The entire process utilizes a parallel pipeline design, keeping single analysis latency within 150 milliseconds.

[0099] The visualization system maps trend indicators into a dynamic radar chart, with three axes representing displacement, strain, and texture trends. In the K25+380 case, the radar chart at 10:15:20 shows significant outward expansion of the strain and texture axes, and a sharp angle on the displacement axis. A heat map also displays the longitudinal risk distribution of the tunnel, with high-risk sections highlighted in dark red. The user interface allows users to click on any data point to view a detailed trend breakdown.

[0100] The exception handling mechanism includes trend conflict detection. When displacement and strain trends reverse (e.g., displacement increases while strain decreases), the system automatically triggers a review process. This review includes checking sensor calibration, re-collecting image data, and manual confirmation. In the K25+380 case, the system detected a mismatch between displacement and texture trends at 10:15:25. Review revealed that this discrepancy in texture analysis was caused by camera lens contamination.

[0101] Data storage utilizes a tiered archiving strategy. Raw trend data is retained for 30 days, feature vectors for 180 days, and risk assessment results are permanently stored. The query interface supports searches based on time range, spatial location, and risk level. The system generates weekly trend analysis reports, including graphs of changes in each indicator and statistical summaries. All analysis results are digitally signed to ensure data integrity and traceability.

[0102] The quality control module regularly performs standard tests. The test dataset contains trend characteristics of 12 typical deformation scenarios, and the system runs the test process every 24 hours. If the detection accuracy deviation exceeds 5%, the model retraining process is automatically triggered. The training process uses an incremental learning approach, incorporating new samples while retaining existing knowledge, and the training time is controlled to complete within 30 minutes. The updated model is verified through shadow runs and then transferred to the production environment.

[0103] Example 5: The monitoring result report adopts a structured data encapsulation format and defines a data model containing six core fields. The tunnel pile number field records the mileage position of the monitoring point, using the kilometer mark + meter mark format accurate to the centimeter level. The timestamp field records the UTC time and appends the time zone information, with an accuracy of milliseconds. The deformation parameter field stores 12 indicators such as three-dimensional displacement, principal strain direction and maximum strain value. The feature recognition result field contains deformation grade classification, crack width estimation and defect distribution coordinates. The trend indicator field stores three floating-point values ​​of displacement trend angle, strain change rate and texture chaos. The risk level field uses an enumeration type to mark the four levels of safety, attention, warning and danger. The report generation module performs data packaging every 10 seconds, uses JSON-LD format to implement semantic description, and compresses the file size to less than 5KB.

[0104] The central monitoring system utilizes a distributed message processing architecture, deploying an Apache Kafka cluster as the data bus. The cluster is configured with three broker nodes, a dedicated topic for report data is set up, and the topic is partitioned into three partitions. The producer API is integrated into the monitoring terminal, pushing report data via a persistent TCP connection. The consumer group consists of three instances: an alarm service, a storage service, and a visualization service, using a collaborative consumption model for load balancing. The message queue is set to retain data for the last 24 hours, and disk storage uses an SSD array to ensure high throughput. The system's peak processing capacity reaches 2,000 reports per second, with average end-to-end latency controlled at 50 milliseconds.

[0105] The alarm triggering mechanism implements a graded response strategy. When the risk level field enters the caution state (corresponding to a value of 30-60), the yellow alert process is triggered: a yellow indicator light illuminates on the control center console, the affected pile number range is marked on the electronic signage, and a lightweight notification is pushed to the mobile device. When the risk value rises to the 60-80 range, an orange alert is triggered: the audible and visual alarm system is activated, a voice prompt plays in a loop, a response plan including support and reinforcement recommendations is automatically generated, and an SMS message is sent to the responsible engineer. A red alert state is entered when the risk value exceeds 80: power to the construction site in the affected area is cut off, an emergency evacuation broadcast is triggered, and the tunnel structure safety assessment model is simultaneously activated. All alert events are recorded with a response timeline accurate to the millisecond.

[0106] The data persistence system utilizes a hybrid storage architecture. Raw monitoring reports are written to the InfluxDB time-series database, partitioned by time and spatially indexed. Analysis results are persisted to a PostgreSQL relational database, including schematized reports and associated metadata. A Redis cluster is deployed in the cache layer to store the most recent 15 minutes of hot data. The archiving system performs monthly cold data migration, dumping historical data to an object storage system with a permanent retention policy. The query interface supports both ODBC and JDBC protocols, providing multi-dimensional search capabilities based on time range, spatial location, and risk level.

[0107] System interface services are exposed through an API gateway, deploying the OAuth2.0 authentication protocol. External access requires an access token, which is valid for two hours. The data query interface supports both JSON and XML formats, with a response time commitment of 99% of requests completed within 300 milliseconds. The management interface provides device status monitoring, alarm history query, and system configuration update functions. Key operations require dual authentication. Interface traffic is rate-limited, with a single client not exceeding 100 requests per second. All transmitted data is encrypted using TLS1.3, with key fields additionally encrypted using AES-256.

[0108] The blockchain evidence storage system builds a private chain network consisting of four consensus nodes. Each time a monitoring report is generated, a SHA-256 hash value is calculated and written to the blockchain as a digital fingerprint. When an early warning event is triggered, a Merkle tree root value is generated and uploaded to the blockchain, along with complete event details, including the response action, operator, and execution time. The audit trail interface supports input of time ranges and operation types, and returns an immutable operation log. Off-chain backup of evidence data is performed every 24 hours, with backup files stored in physically isolated storage.

[0109] The mobile interactive system has been developed as a cross-platform application, supporting both iOS and Android. Warning message push notifications integrate APNs and FCM dual channels to ensure message reachability. The application interface is divided into three display areas: a real-time monitoring view displays a heat map of tunnel longitudinal risk, a historical analysis view provides a playback of data trends for any time period, and an early warning management view lists unhandled alarms. Offline mode caches the last 72 hours of data and automatically synchronizes operation logs upon network restoration. Device binding utilizes a two-way authentication mechanism, allowing remote data erasure in the event of a lost device.

[0110] The system maintenance module enables full lifecycle management. The configuration center supports dynamic parameter updates, with modifications taking effect immediately without requiring a reboot. The logging system collects operation logs from all components and enables centralized analysis using the ELK stack. The monitoring panel displays 50 metrics in real time, including CPU load, memory usage, and network traffic. The self-diagnostic program performs hardware checks, data consistency checks, and performance benchmarks every morning. The maintenance terminal provides a command-line interface, enabling remote troubleshooting and system repair.

[0111] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0112] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A tunnel construction deformation monitoring method based on machine vision, characterized in that: The method comprises: Real-time image acquisition of tunnel construction structures is performed to obtain tunnel surface image sequences while monitoring the displacement and strain parameters of the tunnel structure. According to the displacement parameters and strain parameters, the initial deformation is predicted to obtain the predicted deformation parameters; Using machine vision equipment, images of key areas inside the tunnel are collected. Based on the predicted deformation parameters, these images are input into multiple deformation feature identifiers to identify the actual deformation features. Combining the predicted deformation parameters with the actual deformation characteristics, the final deformation trend analysis is performed to obtain the tunnel deformation monitoring results.

2. The method for monitoring tunnel construction deformation based on machine vision according to claim 1, characterized in that: Real-time image acquisition of tunnel construction structures is performed to obtain tunnel surface image sequences, while monitoring the displacement and strain parameters of the tunnel structure, including: Deploy multiple cameras at key points in the tunnel to continuously capture tunnel surface images and form a tunnel surface image sequence; Use sensor networks to measure real-time displacement changes of tunnel structures and obtain displacement parameters; Collect the strain data of the tunnel material through the strain gauge to obtain the strain parameters; The displacement and strain parameters are associated with the timestamps of the tunnel surface image sequence.

3. The method for monitoring tunnel construction deformation based on machine vision according to claim 1, characterized in that: Based on the displacement parameters and strain parameters, the initial deformation prediction is performed to obtain the predicted deformation parameters, including: Collect historical tunnel construction data, extract sample displacement parameter sets and sample strain parameter sets, and mark the sample deformation dimensions to form a sample deformation parameter set; Build a deformation prediction model based on time series analysis; Using the sample displacement parameter set, sample strain parameter set, and sample deformation parameter set as training data and test data, the deformation prediction model is trained and tested, and the model optimization is completed after the error rate meets the standard; Inputting displacement parameters and strain parameters into the deformation prediction model, and obtaining predicted deformation parameters through prediction output; Compare the predicted deformation parameters with the real-time displacement parameters to calibrate the model deviation; Based on the calibration model deviation, the deformation prediction model parameters are updated.

4. The method for monitoring tunnel construction deformation based on machine vision according to claim 1, characterized in that: Using machine vision equipment, images of key areas inside the tunnel are collected. Based on the predicted deformation parameters, these images are fed into multiple deformation feature identifiers to identify the actual deformation features, including: Use machine vision equipment to focus on high-stress areas in the tunnel and capture images of key areas; According to the predicted deformation parameters, the matching deformation level range is screened; selecting a plurality of deformation feature identifiers corresponding to a deformation level range, each deformation feature identifier including a plurality of feature recognition paths based on image pattern matching; The key area image is input into multiple deformation feature identifiers, and each deformation feature identifier outputs a binary recognition result; Count the proportion of binary recognition results that are yes and calculate the probability distribution of deformation level; The deformation level with the highest probability is selected as the actual deformation feature; According to the actual deformation characteristics, feedback is provided to adjust the acquisition angle of the key area image.

5. The method for monitoring tunnel construction deformation based on machine vision according to claim 4, characterized in that: A plurality of deformation feature identifiers corresponding to a deformation level range are selected, each deformation feature identifier including a plurality of feature recognition paths based on image pattern matching, including: Pre-training multiple deformation feature identifiers corresponding to multiple deformation levels, the training data including sample key area images and sample deformation feature binary results; determining the number of feature recognition paths based on the error margin of the predicted deformation parameters; randomly activating feature recognition paths within a plurality of deformable feature recognizers; Input the key area image into the activated feature recognition path and output a set of binary recognition results; Aggregate the binary recognition result set and calculate the actual deformation features.

6. The method for monitoring tunnel construction deformation based on machine vision according to claim 1, characterized in that: Combining the predicted deformation parameters with the actual deformation characteristics, the final deformation trend analysis is performed to obtain tunnel deformation monitoring results, including: Integrate the predicted deformation parameters and actual deformation characteristics to form a deformation data vector; Based on the moving window technique, the deformed data vector is time-sliced; Analyze the deformation fluctuation characteristics within the time slice and calculate the average change rate of the deformation fluctuation characteristics; Construct deformation trend indicators and generate tunnel deformation monitoring results based on the deformation trend indicators; Compare tunnel deformation monitoring results with historical data to verify accuracy.

7. The method for monitoring tunnel construction deformation based on machine vision according to claim 6, characterized in that: Based on the moving window technology, the deformed data vector is time-sliced, including setting a fixed time window length and a sliding step size, and dividing the deformed data vector into time series.

8. The method for monitoring tunnel construction deformation based on machine vision according to claim 6, characterized in that: Construct deformation trend indicators, including: Analyze the changing direction of deformation data within the time slice and calculate the upward and downward trend amounts of deformation data; Based on the uptrend volume and the downtrend volume, the deformed trend indicator is obtained.

9. The method for monitoring tunnel construction deformation based on machine vision according to claim 1, characterized in that: Also includes: The tunnel deformation monitoring results are formatted into reports, transmitted to the central monitoring system, and an alarm mechanism is triggered based on the tunnel deformation monitoring results.

10. A tunnel construction deformation monitoring system based on machine vision, characterized in that: A system for implementing the machine vision-based tunnel construction deformation monitoring method according to any one of claims 1 to 9, comprising: Image acquisition module, used to collect real-time images of tunnel construction structures, obtain tunnel surface image sequences, and simultaneously monitor the displacement and strain parameters of the tunnel structure; The prediction module is used to perform initial deformation prediction based on displacement parameters and strain parameters to obtain predicted deformation parameters; The feature recognition module is used to collect images of key areas inside the tunnel through machine vision equipment. Based on the predicted deformation parameters, the images of key areas are input into multiple deformation feature identifiers to identify and obtain actual deformation features. The analysis module is used to combine the predicted deformation parameters and actual deformation characteristics to conduct final deformation trend analysis and obtain tunnel deformation monitoring results.

Images